In this notebook, some template code has already been provided for you, and you will need to implement additional functionality to successfully complete this project. You will not need to modify the included code beyond what is requested. Sections that begin with '(IMPLEMENTATION)' in the header indicate that the following block of code will require additional functionality which you must provide. Instructions will be provided for each section, and the specifics of the implementation are marked in the code block with a 'TODO' statement. Please be sure to read the instructions carefully!
Note: Once you have completed all of the code implementations, you need to finalize your work by exporting the iPython Notebook as an HTML document. Before exporting the notebook to html, all of the code cells need to have been run so that reviewers can see the final implementation and output. You can then export the notebook by using the menu above and navigating to \n", "File -> Download as -> HTML (.html). Include the finished document along with this notebook as your submission.
In addition to implementing code, there will be questions that you must answer which relate to the project and your implementation. Each section where you will answer a question is preceded by a 'Question X' header. Carefully read each question and provide thorough answers in the following text boxes that begin with 'Answer:'. Your project submission will be evaluated based on your answers to each of the questions and the implementation you provide.
Note: Code and Markdown cells can be executed using the Shift + Enter keyboard shortcut. Markdown cells can be edited by double-clicking the cell to enter edit mode.
The rubric contains optional "Stand Out Suggestions" for enhancing the project beyond the minimum requirements. If you decide to pursue the "Stand Out Suggestions", you should include the code in this IPython notebook.
In this notebook, you will make the first steps towards developing an algorithm that could be used as part of a mobile or web app. At the end of this project, your code will accept any user-supplied image as input. If a dog is detected in the image, it will provide an estimate of the dog's breed. If a human is detected, it will provide an estimate of the dog breed that is most resembling. The image below displays potential sample output of your finished project (... but we expect that each student's algorithm will behave differently!).

In this real-world setting, you will need to piece together a series of models to perform different tasks; for instance, the algorithm that detects humans in an image will be different from the CNN that infers dog breed. There are many points of possible failure, and no perfect algorithm exists. Your imperfect solution will nonetheless create a fun user experience!
We break the notebook into separate steps. Feel free to use the links below to navigate the notebook.
In the code cell below, we import a dataset of dog images. We populate a few variables through the use of the load_files function from the scikit-learn library:
train_files, valid_files, test_files - numpy arrays containing file paths to imagestrain_targets, valid_targets, test_targets - numpy arrays containing onehot-encoded classification labels dog_names - list of string-valued dog breed names for translating labelsfrom sklearn.datasets import load_files
from keras.utils import np_utils
import numpy as np
from glob import glob
# define function to load train, test, and validation datasets
def load_dataset(path):
data = load_files(path)
dog_files = np.array(data['filenames'])
dog_targets = np_utils.to_categorical(np.array(data['target']), 133)
return dog_files, dog_targets
# load train, test, and validation datasets
train_files, train_targets = load_dataset('dogImages/train')
valid_files, valid_targets = load_dataset('dogImages/valid')
test_files, test_targets = load_dataset('dogImages/test')
# load list of dog names
dog_names = [item[20:-1] for item in sorted(glob("dogImages/train/*/"))]
# print statistics about the dataset
print('There are %d total dog categories.' % len(dog_names))
print('There are %s total dog images.\n' % len(np.hstack([train_files, valid_files, test_files])))
print('There are %d training dog images.' % len(train_files))
print('There are %d validation dog images.' % len(valid_files))
print('There are %d test dog images.'% len(test_files))
There are 133 total dog categories. There are 8351 total dog images. There are 6680 training dog images. There are 835 validation dog images. There are 836 test dog images.
# Since I can't reference pictures directly from the udacity urls,
# we are gonna use a helper function to embed images properly using html.
from IPython.display import Image, HTML, display
import base64
def to_dog_labels(arr):
return dog_names[arr.argmax()].split("/")[-1]
def img_to_html(img_path, figcaption="", height="150px"):
'''convert image to data in base64
then construct html img block
'''
figure_html = '''
<figure style= '{figstyle}'>
<img style='{imgstyle}' src='{imgsrc}'/>
<figcaption style='font-size:80%; text-align:center'>{figcaption}</figcaption>
</figure>
'''
figstyle = "margin: 10px; float: left; border: 3px solid black;"
imgstyle = "height:{}; padding: 5px;".format(height)
img_base64 = str(base64.b64encode(Image(img_path).data))[2:]
html = figure_html.format(
figstyle = figstyle,
imgstyle = imgstyle,
imgsrc = 'data:image/jpeg;base64,' + img_base64,
figcaption = figcaption
)
return html
def gallery(files, labels="", height="150px"):
'''uses html display to show image files inline in jupyter'''
images_list = ""
if not labels :
for file in files:
images_list += (img_to_html(file, figcaption="", height=height))
else :
for file, label in zip(files, labels):
images_list += (img_to_html(file, figcaption=label, height=height))
display(HTML(images_list))
gallery(train_files[:18], [to_dog_labels(t) for t in train_targets[:18]])
In the code cell below, we import a dataset of human images, where the file paths are stored in the numpy array human_files.
import random
random.seed(8675319)
# load filenames in shuffled human dataset
human_files = np.array(glob("lfw/*/*"))
random.shuffle(human_files)
# print statistics about the dataset
print('There are %d total human images.' % len(human_files))
There are 13233 total human images.
We use OpenCV's implementation of Haar feature-based cascade classifiers to detect human faces in images. OpenCV provides many pre-trained face detectors, stored as XML files on github. We have downloaded one of these detectors and stored it in the haarcascades directory.
In the next code cell, we demonstrate how to use this detector to find human faces in a sample image.
import cv2
import matplotlib.pyplot as plt
%matplotlib inline
# extract pre-trained face detector
face_cascade = cv2.CascadeClassifier('haarcascades/haarcascade_frontalface_alt.xml')
# load color (BGR) image
img = cv2.imread(human_files[3])
# convert BGR image to grayscale
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
# find faces in image
faces = face_cascade.detectMultiScale(gray)
# print number of faces detected in the image
print('Number of faces detected:', len(faces))
# get bounding box for each detected face
for (x,y,w,h) in faces:
# add bounding box to color image
cv2.rectangle(img,(x,y),(x+w,y+h),(255,0,0),2)
# convert BGR image to RGB for plotting
cv_rgb = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
# display the image, along with bounding box
plt.imshow(cv_rgb)
plt.show()
Number of faces detected: 1
Before using any of the face detectors, it is standard procedure to convert the images to grayscale. The detectMultiScale function executes the classifier stored in face_cascade and takes the grayscale image as a parameter.
In the above code, faces is a numpy array of detected faces, where each row corresponds to a detected face. Each detected face is a 1D array with four entries that specifies the bounding box of the detected face. The first two entries in the array (extracted in the above code as x and y) specify the horizontal and vertical positions of the top left corner of the bounding box. The last two entries in the array (extracted here as w and h) specify the width and height of the box.
We can use this procedure to write a function that returns True if a human face is detected in an image and False otherwise. This function, aptly named face_detector, takes a string-valued file path to an image as input and appears in the code block below.
# returns "True" if face is detected in image stored at img_path
def face_detector(img_path):
img = cv2.imread(img_path)
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
faces = face_cascade.detectMultiScale(gray)
return len(faces) > 0
Question 1: Use the code cell below to test the performance of the face_detector function.
human_files have a detected human face? dog_files have a detected human face? Ideally, we would like 100% of human images with a detected face and 0% of dog images with a detected face. You will see that our algorithm falls short of this goal, but still gives acceptable performance. We extract the file paths for the first 100 images from each of the datasets and store them in the numpy arrays human_files_short and dog_files_short.
Answer:
human_files_short = human_files[:100]
dog_files_short = train_files[:100]
# Do NOT modify the code above this line.
## TODO: Test the performance of the face_detector algorithm
## on the images in human_files_short and dog_files_short.
# small helper function
def faces_positive(img_list):
'''implements face_detector over an image list.
Returns the sum of positive outcomes in a list.
'''
results = []
for img in img_list:
results.append(face_detector(img))
return np.sum(results)
print("What percentage of the first 100 images in human_files have a detected human face?")
print((faces_positive(human_files_short)/len(human_files_short))*100)
print("What percentage of the first 100 images in dog_files have a detected human face?")
print((faces_positive(dog_files_short)/len(dog_files_short))*100)
What percentage of the first 100 images in human_files have a detected human face? 97.0 What percentage of the first 100 images in dog_files have a detected human face? 12.0
Question 2: This algorithmic choice necessitates that we communicate to the user that we accept human images only when they provide a clear view of a face (otherwise, we risk having unneccessarily frustrated users!). In your opinion, is this a reasonable expectation to pose on the user? If not, can you think of a way to detect humans in images that does not necessitate an image with a clearly presented face?
Answer:
We suggest the face detector from OpenCV as a potential way to detect human images in your algorithm, but you are free to explore other approaches, especially approaches that make use of deep learning :). Please use the code cell below to design and test your own face detection algorithm. If you decide to pursue this optional task, report performance on each of the datasets.
## (Optional) TODO: Report the performance of another
## face detection algorithm on the LFW dataset
### Feel free to use as many code cells as needed.
import dlib
from imutils import face_utils
# Original human face detection Haar-like
def face_detector(img_path):
"""Returns True if a face is found."""
img = cv2.imread(img_path)
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
faces = face_cascade.detectMultiScale(gray)
return len(faces) > 0
# Alternative algorithm # 1
# Histogram of Oriented Gradients (HOG) in Dlib
# https://lear.inrialpes.fr/people/triggs/pubs/Dalal-cvpr05.pdf
def hog_face_detector(img_path):
"""Returns True if a face is found."""
face_hog = dlib.get_frontal_face_detector()
img = cv2.imread(img_path)
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
rects = face_hog(gray, 1)
return len(rects) > 0
# Alternative algorithm # 2
# Convolutional Neural Network in Dlib
# model file obtained here: http://dlib.net/files/mmod_human_face_detector.dat.bz2
def cnn_face_detector(img_path):
"""Returns True if a face is found."""
face_cnn = dlib.cnn_face_detection_model_v1("mmod/mmod_human_face_detector.dat")
img = cv2.imread(img_path)
# Resize the image to attain reasonable speed:
# img = cv2.resize(img, (0, 0), fx=0.5, fy=0.5)
rects = face_cnn(img, 1)
return len(rects) > 0
# Creating a loop function to compare Haar-like vs 2 alternative models.
from tqdm import tqdm
def compare_human_detectors(test_images, detectors):
'''Takes a list of images paths and test them against a list of detectors,
comparing and printig the final detection rate.
Returns a list of dictionaries containing results.
Note : detectors must be functions returning a boolean.
'''
results = []
size = len(test_images)
print("Begin test of sample of ", test_images.shape[0])
for img in tqdm(test_images):
d = {
"img_path":None,
"results" : [],
}
d["img_path"] = img
# Test against each detector
for i, detector in enumerate(detectors):
d["results"].append(detector(img))
results.append(d)
# bar1.update(int(1))
# Print detection rate
for i in range(len(detectors)):
detection_rate = np.sum([x["results"][i] for x in results]) / size * 100
print(" detector # {} : {:.2f} %".format(i+1, detection_rate))
return results
# Run test 1: detect humans in humans
# ⚠️ this may take a while because of CNN-MMOD's computations times
detectors = [
face_detector,
hog_face_detector,
cnn_face_detector
]
print("Part 1: humans in humans")
hh = compare_human_detectors(human_files_short, detectors=detectors)
Part 1: humans in humans Begin test of sample of 100
100%|██████████| 100/100 [05:38<00:00, 3.39s/it]
detector # 1 : 97.00 % detector # 2 : 99.00 % detector # 3 : 100.00 %
# Run test 1: humans in dog images
# ⚠️ this may take a while because of CNN-MMOD's computations times
print("Part 2: humans in dogs")
hd = compare_human_detectors(dog_files_short, detectors=detectors)
Part 2: humans in dogs Begin test of sample of 100
100%|██████████| 100/100 [34:26<00:00, 20.66s/it]
detector # 1 : 12.00 % detector # 2 : 8.00 % detector # 3 : 2.00 %
#that took a while. I'm gonna save those results
import json
# with open("hh_results.json", "w") as jp:
# json.dump(hh, jp)
# with open("hd_results.json", "w") as jp:
# json.dump(hd, jp)
# relading
# Opening JSON file
with open('hh_results.json') as json_file:
hh = json.load(json_file)
with open('hd_results.json') as json_file:
hd = json.load(json_file)
# take a look at those inconsistencies in human detection
arr = np.sum([i["results"] for i in hh], axis=1)
ixs = np.where(arr<3)
print("Showing images with detection discrepancies among models")
images = [hh[ix]["img_path"] for ix in ixs[0]]
labels = [hh[ix]["img_path"].split("\\")[-1] + "</br>" + str(hh[ix]["results"]) for ix in ixs[0]]
gallery(images, labels)
Showing images with detection discrepancies among models
# take a look at those inconsistencies in human in dogs detection
arr = np.sum([i["results"] for i in hd], axis=1)
ixs = np.where(arr>0)
print("Showing images with detection discrepancies among models, for the dogs sample")
images = [hd[ix]["img_path"] for ix in ixs[0]]
labels = [hd[ix]["img_path"].split("\\")[-1] + "</br>" + str(hd[ix]["results"]) for ix in ixs[0]]
gallery(images, labels)
Showing images with detection discrepancies among models, for the dogs sample
from io import BytesIO
import base64
import PIL
import imutils
# drawing boxes for better inspection
def draw_bow(faces, image_np, text, color):
'''draw boxes using faces coordinates and annotates text in image'''
for (x,y,w,h) in faces:
cv2.rectangle(image_np,(x,y),(x+w,y+h), color,2)
cv2.putText(image_np, text, (x,y-3), fontFace = cv2.FONT_HERSHEY_PLAIN, fontScale = 1, color = color)
def bboxes(img_path):
'''implement various face detection models and draw boxes.
Returns image in a 4-dimensions array (numpy.array).'''
img = cv2.imread(img_path)
# resize very large images
if img.shape[0] > 700:
img = imutils.resize(img, width=700)
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
# Haar-like
face_cascade = cv2.CascadeClassifier('haarcascades/haarcascade_frontalface_alt.xml')
faces = face_cascade.detectMultiScale(gray)
if len(faces) > 0:
draw_bow(faces, img, "haar", (255,0,255))
# HoG
face_hog = dlib.get_frontal_face_detector()
rects = face_hog(gray, 1)
if rects :
faces = []
for d in rects :
faces.append(face_utils.rect_to_bb(d))
faces = np.array(faces)
draw_bow(faces, img, "hog-svm", (255,255,0))
# CNN
face_cnn = dlib.cnn_face_detection_model_v1("mmod/mmod_human_face_detector.dat")
rects = face_cnn(img, 1)
if rects:
faces = np.array([face_utils.rect_to_bb(rects[0].rect)])
draw_bow(faces, img, "cnn", (0,255,255))
return cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
def to_b64(image_array):
'''transforms nd.array image into b64 string'''
pil_img = PIL.Image.fromarray(image_array)
buff = BytesIO()
pil_img.save(buff, format="PNG")
b64str = base64.b64encode(buff.getvalue()).decode("utf-8")
return b64str
# Run to see some false positives examples.
# ⚠️ this may take a while because of CNN-MMOD's computations times
images_html = ""
for img_path in tqdm(images):
image_with_bboxes = bboxes(img_path)
figure_html= "<figure style='height: 300px; margin: 20px 10px; float: left'><img src='{imgsrc}'/></figure>"
html = figure_html.format(
imgsrc = 'data:image/jpeg;base64,' + to_b64(image_with_bboxes),
)
images_html += html
display(HTML(images_html))
100%|██████████████████████████████████████████████████████████████████████████████████| 17/17 [03:45<00:00, 13.26s/it]
In this section, we use a pre-trained ResNet-50 model to detect dogs in images. Our first line of code downloads the ResNet-50 model, along with weights that have been trained on ImageNet, a very large, very popular dataset used for image classification and other vision tasks. ImageNet contains over 10 million URLs, each linking to an image containing an object from one of 1000 categories. Given an image, this pre-trained ResNet-50 model returns a prediction (derived from the available categories in ImageNet) for the object that is contained in the image.
from keras.applications.resnet import ResNet50
# define ResNet50 model
ResNet50_model = ResNet50(weights='imagenet')
When using TensorFlow as backend, Keras CNNs require a 4D array (which we'll also refer to as a 4D tensor) as input, with shape
$$ (\text{nb-samples}, \text{rows}, \text{columns}, \text{channels}) $$where nb_samples corresponds to the total number of images (or samples), and rows, columns, and channels correspond to the number of rows, columns, and channels for each image, respectively.
The path_to_tensor function below takes a string-valued file path to a color image as input and returns a 4D tensor suitable for supplying to a Keras CNN. The function first loads the image and resizes it to a square image that is $224 \times 224$ pixels. Next, the image is converted to an array, which is then resized to a 4D tensor. In this case, since we are working with color images, each image has three channels. Likewise, since we are processing a single image (or sample), the returned tensor will always have shape
The paths_to_tensor function takes a numpy array of string-valued image paths as input and returns a 4D tensor with shape
Here, nb_samples is the number of samples, or number of images, in the supplied array of image paths. It is best to think of nb_samples as the number of 3D tensors (where each 3D tensor corresponds to a different image) in your dataset!
import keras
# from tqdm import tqdm
def path_to_tensor(img_path):
# loads RGB image as PIL.Image.Image type
img = keras.utils.load_img(img_path, target_size=(224, 224))
# convert PIL.Image.Image type to 3D tensor with shape (224, 224, 3)
x = keras.utils.img_to_array(img)
# convert 3D tensor to 4D tensor with shape (1, 224, 224, 3) and return 4D tensor
return np.expand_dims(x, axis=0)
def paths_to_tensor(img_paths):
list_of_tensors = [path_to_tensor(img_path) for img_path in tqdm(img_paths)]
return np.vstack(list_of_tensors)
path_to_tensor(human_files[0]).shape
(1, 224, 224, 3)
Getting the 4D tensor ready for ResNet-50, and for any other pre-trained model in Keras, requires some additional processing. First, the RGB image is converted to BGR by reordering the channels. All pre-trained models have the additional normalization step that the mean pixel (expressed in RGB as $[103.939, 116.779, 123.68]$ and calculated from all pixels in all images in ImageNet) must be subtracted from every pixel in each image. This is implemented in the imported function preprocess_input. If you're curious, you can check the code for preprocess_input here.
Now that we have a way to format our image for supplying to ResNet-50, we are now ready to use the model to extract the predictions. This is accomplished with the predict method, which returns an array whose $i$-th entry is the model's predicted probability that the image belongs to the $i$-th ImageNet category. This is implemented in the ResNet50_predict_labels function below.
By taking the argmax of the predicted probability vector, we obtain an integer corresponding to the model's predicted object class, which we can identify with an object category through the use of this dictionary.
from keras.applications.resnet import ResNet50, preprocess_input, decode_predictions
def ResNet50_predict_labels(img_path):
# returns prediction vector for image located at img_path
img = preprocess_input(path_to_tensor(img_path))
return np.argmax(ResNet50_model.predict(img, verbose=0)) # added verbose=0 for silent
While looking at the dictionary, you will notice that the categories corresponding to dogs appear in an uninterrupted sequence and correspond to dictionary keys 151-268, inclusive, to include all categories from 'Chihuahua' to 'Mexican hairless'. Thus, in order to check to see if an image is predicted to contain a dog by the pre-trained ResNet-50 model, we need only check if the ResNet50_predict_labels function above returns a value between 151 and 268 (inclusive).
We use these ideas to complete the dog_detector function below, which returns True if a dog is detected in an image (and False if not).
### returns "True" if a dog is detected in the image stored at img_path
def dog_detector(img_path):
prediction = ResNet50_predict_labels(img_path)
return ((prediction <= 268) & (prediction >= 151))
Question 3: Use the code cell below to test the performance of your dog_detector function.
human_files_short have a detected dog? dog_files_short have a detected dog?Answer:
### TODO: Test the performance of the dog_detector function
### on the images in human_files_short and dog_files_short.
# Loop and store results in list
print("running assessment(s)")
results_dogs_in_humans =[]
results_dogs_in_dogs = []
for img in tqdm(human_files_short):
results_dogs_in_humans.append(dog_detector(img))
for img in tqdm(dog_files_short):
results_dogs_in_dogs.append(dog_detector(img))
# Print results
print("Percentage of dog-detection in human_files_short :")
print((np.sum(results_dogs_in_humans)/len(human_files_short))*100)
print("Percentage of dog-detection in dog_files_short :")
print((np.sum(results_dogs_in_dogs)/len(dog_files_short))*100)
running assessment(s)
100%|████████████████████████████████████████████████████████████████████████████████| 100/100 [00:10<00:00, 9.15it/s] 100%|████████████████████████████████████████████████████████████████████████████████| 100/100 [00:11<00:00, 8.96it/s]
Percentage of dog-detection in human_files_short : 0.0 Percentage of dog-detection in dog_files_short : 100.0
Now that we have functions for detecting humans and dogs in images, we need a way to predict breed from images. In this step, you will create a CNN that classifies dog breeds. You must create your CNN from scratch (so, you can't use transfer learning yet!), and you must attain a test accuracy of at least 1%. In Step 5 of this notebook, you will have the opportunity to use transfer learning to create a CNN that attains greatly improved accuracy.
Be careful with adding too many trainable layers! More parameters means longer training, which means you are more likely to need a GPU to accelerate the training process. Thankfully, Keras provides a handy estimate of the time that each epoch is likely to take; you can extrapolate this estimate to figure out how long it will take for your algorithm to train.
We mention that the task of assigning breed to dogs from images is considered exceptionally challenging. To see why, consider that even a human would have great difficulty in distinguishing between a Brittany and a Welsh Springer Spaniel.
| Brittany | Welsh Springer Spaniel |
|---|---|
![]() |
![]() |
It is not difficult to find other dog breed pairs with minimal inter-class variation (for instance, Curly-Coated Retrievers and American Water Spaniels).
| Curly-Coated Retriever | American Water Spaniel |
|---|---|
![]() |
![]() |
Likewise, recall that labradors come in yellow, chocolate, and black. Your vision-based algorithm will have to conquer this high intra-class variation to determine how to classify all of these different shades as the same breed.
| Yellow Labrador | Chocolate Labrador | Black Labrador |
|---|---|---|
![]() |
![]() |
![]() |
We also mention that random chance presents an exceptionally low bar: setting aside the fact that the classes are slightly imabalanced, a random guess will provide a correct answer roughly 1 in 133 times, which corresponds to an accuracy of less than 1%.
Remember that the practice is far ahead of the theory in deep learning. Experiment with many different architectures, and trust your intuition. And, of course, have fun!
We rescale the images by dividing every pixel in every image by 255.
from PIL import ImageFile
ImageFile.LOAD_TRUNCATED_IMAGES = True
from tqdm import tqdm
# pre-process the data for Keras
train_tensors = paths_to_tensor(train_files).astype('float32')/255
valid_tensors = paths_to_tensor(valid_files).astype('float32')/255
test_tensors = paths_to_tensor(test_files).astype('float32')/255
100%|█████████████████████████████████████████████████████████████████████████████| 6680/6680 [00:31<00:00, 212.98it/s] 100%|███████████████████████████████████████████████████████████████████████████████| 835/835 [00:03<00:00, 224.13it/s] 100%|███████████████████████████████████████████████████████████████████████████████| 836/836 [00:03<00:00, 231.37it/s]
Create a CNN to classify dog breed. At the end of your code cell block, summarize the layers of your model by executing the line:
model.summary()
We have imported some Python modules to get you started, but feel free to import as many modules as you need. If you end up getting stuck, here's a hint that specifies a model that trains relatively fast on CPU and attains >1% test accuracy in 5 epochs:

Question 4: Outline the steps you took to get to your final CNN architecture and your reasoning at each step. If you chose to use the hinted architecture above, describe why you think that CNN architecture should work well for the image classification task.
Answer:
import keras
from tensorflow.keras import layers
### TODO: Define your architecture.
inputs = keras.Input(shape=((224, 224, 3)))
# best one overall
model = keras.Sequential(
[
inputs,
layers.Conv2D(16, kernel_size=3, activation='relu'),
layers.MaxPooling2D(pool_size=3, strides=2),
layers.Conv2D(32, kernel_size=3, activation='relu', padding="same"),
layers.MaxPooling2D(pool_size=3, strides=2),
layers.Conv2D(16, kernel_size=3, activation='relu', padding="same"),
layers.MaxPooling2D(pool_size=3, strides=2),
layers.GlobalAveragePooling2D(),
layers.Dense(units=133, activation="softmax") #sigmoid or softmax are suited for multiple classes
]
)
model.summary(line_length=90)
Model: "sequential_1"
__________________________________________________________________________________________
Layer (type) Output Shape Param #
==========================================================================================
conv2d (Conv2D) (None, 222, 222, 16) 448
max_pooling2d (MaxPooling2D) (None, 110, 110, 16) 0
conv2d_1 (Conv2D) (None, 110, 110, 32) 4640
max_pooling2d_1 (MaxPooling2D) (None, 54, 54, 32) 0
conv2d_2 (Conv2D) (None, 54, 54, 16) 4624
max_pooling2d_2 (MaxPooling2D) (None, 26, 26, 16) 0
global_average_pooling2d (GlobalAverag (None, 16) 0
ePooling2D)
dense (Dense) (None, 133) 2261
==========================================================================================
Total params: 11,973
Trainable params: 11,973
Non-trainable params: 0
__________________________________________________________________________________________
model.compile(optimizer='rmsprop', loss='categorical_crossentropy', metrics=['accuracy'])
Train your model in the code cell below. Use model checkpointing to save the model that attains the best validation loss.
You are welcome to augment the training data, but this is not a requirement.
import datetime
### TODO: specify the number of epochs that you would like to use to train the model.
epochs = 50
### Do NOT modify the code below this line.
log_dir = "logs/fit/" + datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
callbacks = [
keras.callbacks.ModelCheckpoint(
filepath='saved_models/weights.best.from_scratch.hdf5',
verbose=1,
save_best_only=True
),
keras.callbacks.TensorBoard(
log_dir = log_dir
)
]
hist = model.fit(
train_tensors,
train_targets,
validation_data=(valid_tensors, valid_targets),
epochs=epochs,
batch_size=20,
callbacks=callbacks,
verbose=1
)
Epoch 1/50 334/334 [==============================] - ETA: 0s - loss: 4.8816 - accuracy: 0.0103 Epoch 1: val_loss improved from inf to 4.86501, saving model to saved_models\weights.best.from_scratch.hdf5 334/334 [==============================] - 30s 87ms/step - loss: 4.8816 - accuracy: 0.0103 - val_loss: 4.8650 - val_accuracy: 0.0180 Epoch 2/50 334/334 [==============================] - ETA: 0s - loss: 4.8433 - accuracy: 0.0129 Epoch 2: val_loss improved from 4.86501 to 4.82942, saving model to saved_models\weights.best.from_scratch.hdf5 334/334 [==============================] - 35s 105ms/step - loss: 4.8433 - accuracy: 0.0129 - val_loss: 4.8294 - val_accuracy: 0.0204 Epoch 3/50 334/334 [==============================] - ETA: 0s - loss: 4.8010 - accuracy: 0.0153 Epoch 3: val_loss improved from 4.82942 to 4.78743, saving model to saved_models\weights.best.from_scratch.hdf5 334/334 [==============================] - 32s 95ms/step - loss: 4.8010 - accuracy: 0.0153 - val_loss: 4.7874 - val_accuracy: 0.0287 Epoch 4/50 334/334 [==============================] - ETA: 0s - loss: 4.7560 - accuracy: 0.0187 Epoch 4: val_loss improved from 4.78743 to 4.75643, saving model to saved_models\weights.best.from_scratch.hdf5 334/334 [==============================] - 32s 97ms/step - loss: 4.7560 - accuracy: 0.0187 - val_loss: 4.7564 - val_accuracy: 0.0216 Epoch 5/50 334/334 [==============================] - ETA: 0s - loss: 4.7231 - accuracy: 0.0234 Epoch 5: val_loss improved from 4.75643 to 4.73187, saving model to saved_models\weights.best.from_scratch.hdf5 334/334 [==============================] - 33s 97ms/step - loss: 4.7231 - accuracy: 0.0234 - val_loss: 4.7319 - val_accuracy: 0.0216 Epoch 6/50 334/334 [==============================] - ETA: 0s - loss: 4.6935 - accuracy: 0.0257 Epoch 6: val_loss improved from 4.73187 to 4.71594, saving model to saved_models\weights.best.from_scratch.hdf5 334/334 [==============================] - 28s 82ms/step - loss: 4.6935 - accuracy: 0.0257 - val_loss: 4.7159 - val_accuracy: 0.0228 Epoch 7/50 334/334 [==============================] - ETA: 0s - loss: 4.6637 - accuracy: 0.0295 Epoch 7: val_loss improved from 4.71594 to 4.69795, saving model to saved_models\weights.best.from_scratch.hdf5 334/334 [==============================] - 27s 82ms/step - loss: 4.6637 - accuracy: 0.0295 - val_loss: 4.6979 - val_accuracy: 0.0251 Epoch 8/50 334/334 [==============================] - ETA: 0s - loss: 4.6344 - accuracy: 0.0352 Epoch 8: val_loss improved from 4.69795 to 4.69542, saving model to saved_models\weights.best.from_scratch.hdf5 334/334 [==============================] - 29s 86ms/step - loss: 4.6344 - accuracy: 0.0352 - val_loss: 4.6954 - val_accuracy: 0.0228 Epoch 9/50 334/334 [==============================] - ETA: 0s - loss: 4.6063 - accuracy: 0.0388 Epoch 9: val_loss improved from 4.69542 to 4.66292, saving model to saved_models\weights.best.from_scratch.hdf5 334/334 [==============================] - 26s 78ms/step - loss: 4.6063 - accuracy: 0.0388 - val_loss: 4.6629 - val_accuracy: 0.0323 Epoch 10/50 334/334 [==============================] - ETA: 0s - loss: 4.5766 - accuracy: 0.0388 Epoch 10: val_loss improved from 4.66292 to 4.64510, saving model to saved_models\weights.best.from_scratch.hdf5 334/334 [==============================] - 26s 78ms/step - loss: 4.5766 - accuracy: 0.0388 - val_loss: 4.6451 - val_accuracy: 0.0299 Epoch 11/50 334/334 [==============================] - ETA: 0s - loss: 4.5453 - accuracy: 0.0442 Epoch 11: val_loss improved from 4.64510 to 4.63822, saving model to saved_models\weights.best.from_scratch.hdf5 334/334 [==============================] - 30s 89ms/step - loss: 4.5453 - accuracy: 0.0442 - val_loss: 4.6382 - val_accuracy: 0.0395 Epoch 12/50 334/334 [==============================] - ETA: 0s - loss: 4.5211 - accuracy: 0.0433 Epoch 12: val_loss improved from 4.63822 to 4.59973, saving model to saved_models\weights.best.from_scratch.hdf5 334/334 [==============================] - 29s 87ms/step - loss: 4.5211 - accuracy: 0.0433 - val_loss: 4.5997 - val_accuracy: 0.0395 Epoch 13/50 334/334 [==============================] - ETA: 0s - loss: 4.4861 - accuracy: 0.0455 Epoch 13: val_loss improved from 4.59973 to 4.57351, saving model to saved_models\weights.best.from_scratch.hdf5 334/334 [==============================] - 26s 77ms/step - loss: 4.4861 - accuracy: 0.0455 - val_loss: 4.5735 - val_accuracy: 0.0359 Epoch 14/50 334/334 [==============================] - ETA: 0s - loss: 4.4599 - accuracy: 0.0540 Epoch 14: val_loss improved from 4.57351 to 4.53750, saving model to saved_models\weights.best.from_scratch.hdf5 334/334 [==============================] - 26s 76ms/step - loss: 4.4599 - accuracy: 0.0540 - val_loss: 4.5375 - val_accuracy: 0.0335 Epoch 15/50 334/334 [==============================] - ETA: 0s - loss: 4.4230 - accuracy: 0.0557 Epoch 15: val_loss did not improve from 4.53750 334/334 [==============================] - 31s 93ms/step - loss: 4.4230 - accuracy: 0.0557 - val_loss: 4.6045 - val_accuracy: 0.0287 Epoch 16/50 334/334 [==============================] - ETA: 0s - loss: 4.3931 - accuracy: 0.0612 Epoch 16: val_loss improved from 4.53750 to 4.52353, saving model to saved_models\weights.best.from_scratch.hdf5 334/334 [==============================] - 26s 78ms/step - loss: 4.3931 - accuracy: 0.0612 - val_loss: 4.5235 - val_accuracy: 0.0263 Epoch 17/50 334/334 [==============================] - ETA: 0s - loss: 4.3615 - accuracy: 0.0594 Epoch 17: val_loss improved from 4.52353 to 4.48902, saving model to saved_models\weights.best.from_scratch.hdf5 334/334 [==============================] - 26s 78ms/step - loss: 4.3615 - accuracy: 0.0594 - val_loss: 4.4890 - val_accuracy: 0.0431 Epoch 18/50 334/334 [==============================] - ETA: 0s - loss: 4.3260 - accuracy: 0.0653 Epoch 18: val_loss improved from 4.48902 to 4.42274, saving model to saved_models\weights.best.from_scratch.hdf5 334/334 [==============================] - 26s 79ms/step - loss: 4.3260 - accuracy: 0.0653 - val_loss: 4.4227 - val_accuracy: 0.0407 Epoch 19/50 334/334 [==============================] - ETA: 0s - loss: 4.2923 - accuracy: 0.0671 Epoch 19: val_loss did not improve from 4.42274 334/334 [==============================] - 26s 78ms/step - loss: 4.2923 - accuracy: 0.0671 - val_loss: 4.4260 - val_accuracy: 0.0527 Epoch 20/50 334/334 [==============================] - ETA: 0s - loss: 4.2660 - accuracy: 0.0754 Epoch 20: val_loss improved from 4.42274 to 4.35517, saving model to saved_models\weights.best.from_scratch.hdf5 334/334 [==============================] - 28s 84ms/step - loss: 4.2660 - accuracy: 0.0754 - val_loss: 4.3552 - val_accuracy: 0.0491 Epoch 21/50 334/334 [==============================] - ETA: 0s - loss: 4.2356 - accuracy: 0.0726 Epoch 21: val_loss did not improve from 4.35517 334/334 [==============================] - 27s 81ms/step - loss: 4.2356 - accuracy: 0.0726 - val_loss: 4.3598 - val_accuracy: 0.0479 Epoch 22/50 334/334 [==============================] - ETA: 0s - loss: 4.2083 - accuracy: 0.0754 Epoch 22: val_loss did not improve from 4.35517 334/334 [==============================] - 31s 92ms/step - loss: 4.2083 - accuracy: 0.0754 - val_loss: 4.3603 - val_accuracy: 0.0575 Epoch 23/50 334/334 [==============================] - ETA: 0s - loss: 4.1877 - accuracy: 0.0756 Epoch 23: val_loss improved from 4.35517 to 4.30869, saving model to saved_models\weights.best.from_scratch.hdf5 334/334 [==============================] - 26s 79ms/step - loss: 4.1877 - accuracy: 0.0756 - val_loss: 4.3087 - val_accuracy: 0.0587 Epoch 24/50 334/334 [==============================] - ETA: 0s - loss: 4.1641 - accuracy: 0.0858 Epoch 24: val_loss improved from 4.30869 to 4.30740, saving model to saved_models\weights.best.from_scratch.hdf5 334/334 [==============================] - 27s 81ms/step - loss: 4.1641 - accuracy: 0.0858 - val_loss: 4.3074 - val_accuracy: 0.0599 Epoch 25/50 334/334 [==============================] - ETA: 0s - loss: 4.1427 - accuracy: 0.0835 Epoch 25: val_loss improved from 4.30740 to 4.25449, saving model to saved_models\weights.best.from_scratch.hdf5 334/334 [==============================] - 28s 83ms/step - loss: 4.1427 - accuracy: 0.0835 - val_loss: 4.2545 - val_accuracy: 0.0587 Epoch 26/50 334/334 [==============================] - ETA: 0s - loss: 4.1206 - accuracy: 0.0871 Epoch 26: val_loss did not improve from 4.25449 334/334 [==============================] - 26s 78ms/step - loss: 4.1206 - accuracy: 0.0871 - val_loss: 4.3178 - val_accuracy: 0.0647 Epoch 27/50 334/334 [==============================] - ETA: 0s - loss: 4.0976 - accuracy: 0.0898 Epoch 27: val_loss did not improve from 4.25449 334/334 [==============================] - 28s 84ms/step - loss: 4.0976 - accuracy: 0.0898 - val_loss: 4.2963 - val_accuracy: 0.0635 Epoch 28/50 334/334 [==============================] - ETA: 0s - loss: 4.0744 - accuracy: 0.0927 Epoch 28: val_loss did not improve from 4.25449 334/334 [==============================] - 26s 77ms/step - loss: 4.0744 - accuracy: 0.0927 - val_loss: 4.2782 - val_accuracy: 0.0611 Epoch 29/50 334/334 [==============================] - ETA: 0s - loss: 4.0502 - accuracy: 0.0954 Epoch 29: val_loss improved from 4.25449 to 4.20806, saving model to saved_models\weights.best.from_scratch.hdf5 334/334 [==============================] - 26s 76ms/step - loss: 4.0502 - accuracy: 0.0954 - val_loss: 4.2081 - val_accuracy: 0.0802 Epoch 30/50 334/334 [==============================] - ETA: 0s - loss: 4.0406 - accuracy: 0.0949 Epoch 30: val_loss improved from 4.20806 to 4.19541, saving model to saved_models\weights.best.from_scratch.hdf5 334/334 [==============================] - 26s 77ms/step - loss: 4.0406 - accuracy: 0.0949 - val_loss: 4.1954 - val_accuracy: 0.0683 Epoch 31/50 334/334 [==============================] - ETA: 0s - loss: 4.0197 - accuracy: 0.0994 Epoch 31: val_loss did not improve from 4.19541 334/334 [==============================] - 26s 77ms/step - loss: 4.0197 - accuracy: 0.0994 - val_loss: 4.2731 - val_accuracy: 0.0623 Epoch 32/50 334/334 [==============================] - ETA: 0s - loss: 3.9935 - accuracy: 0.0996 Epoch 32: val_loss improved from 4.19541 to 4.15689, saving model to saved_models\weights.best.from_scratch.hdf5 334/334 [==============================] - 28s 83ms/step - loss: 3.9935 - accuracy: 0.0996 - val_loss: 4.1569 - val_accuracy: 0.0814 Epoch 33/50 334/334 [==============================] - ETA: 0s - loss: 3.9803 - accuracy: 0.1034 Epoch 33: val_loss did not improve from 4.15689 334/334 [==============================] - 33s 98ms/step - loss: 3.9803 - accuracy: 0.1034 - val_loss: 4.1690 - val_accuracy: 0.0719 Epoch 34/50 334/334 [==============================] - ETA: 0s - loss: 3.9628 - accuracy: 0.1078 Epoch 34: val_loss did not improve from 4.15689 334/334 [==============================] - 32s 95ms/step - loss: 3.9628 - accuracy: 0.1078 - val_loss: 4.1920 - val_accuracy: 0.0862 Epoch 35/50 334/334 [==============================] - ETA: 0s - loss: 3.9474 - accuracy: 0.1087 Epoch 35: val_loss improved from 4.15689 to 4.13415, saving model to saved_models\weights.best.from_scratch.hdf5 334/334 [==============================] - 33s 100ms/step - loss: 3.9474 - accuracy: 0.1087 - val_loss: 4.1341 - val_accuracy: 0.0683 Epoch 36/50 334/334 [==============================] - ETA: 0s - loss: 3.9332 - accuracy: 0.1091 Epoch 36: val_loss did not improve from 4.13415 334/334 [==============================] - 34s 100ms/step - loss: 3.9332 - accuracy: 0.1091 - val_loss: 4.2382 - val_accuracy: 0.0659 Epoch 37/50 334/334 [==============================] - ETA: 0s - loss: 3.9212 - accuracy: 0.1078 Epoch 37: val_loss improved from 4.13415 to 4.05745, saving model to saved_models\weights.best.from_scratch.hdf5 334/334 [==============================] - 26s 78ms/step - loss: 3.9212 - accuracy: 0.1078 - val_loss: 4.0575 - val_accuracy: 0.0922 Epoch 38/50 334/334 [==============================] - ETA: 0s - loss: 3.9004 - accuracy: 0.1105 Epoch 38: val_loss did not improve from 4.05745 334/334 [==============================] - 26s 79ms/step - loss: 3.9004 - accuracy: 0.1105 - val_loss: 4.0759 - val_accuracy: 0.0826 Epoch 39/50 334/334 [==============================] - ETA: 0s - loss: 3.8902 - accuracy: 0.1183 Epoch 39: val_loss did not improve from 4.05745 334/334 [==============================] - 28s 84ms/step - loss: 3.8902 - accuracy: 0.1183 - val_loss: 4.1000 - val_accuracy: 0.0766 Epoch 40/50 334/334 [==============================] - ETA: 0s - loss: 3.8772 - accuracy: 0.1177 Epoch 40: val_loss did not improve from 4.05745 334/334 [==============================] - 26s 78ms/step - loss: 3.8772 - accuracy: 0.1177 - val_loss: 4.0726 - val_accuracy: 0.0826 Epoch 41/50 334/334 [==============================] - ETA: 0s - loss: 3.8599 - accuracy: 0.1199 Epoch 41: val_loss did not improve from 4.05745 334/334 [==============================] - 26s 79ms/step - loss: 3.8599 - accuracy: 0.1199 - val_loss: 4.1577 - val_accuracy: 0.0898 Epoch 42/50 334/334 [==============================] - ETA: 0s - loss: 3.8494 - accuracy: 0.1178 Epoch 42: val_loss improved from 4.05745 to 4.03318, saving model to saved_models\weights.best.from_scratch.hdf5 334/334 [==============================] - 26s 79ms/step - loss: 3.8494 - accuracy: 0.1178 - val_loss: 4.0332 - val_accuracy: 0.0874 Epoch 43/50 334/334 [==============================] - ETA: 0s - loss: 3.8385 - accuracy: 0.1244 Epoch 43: val_loss did not improve from 4.03318 334/334 [==============================] - 31s 92ms/step - loss: 3.8385 - accuracy: 0.1244 - val_loss: 4.1212 - val_accuracy: 0.0970 Epoch 44/50 334/334 [==============================] - ETA: 0s - loss: 3.8205 - accuracy: 0.1237 Epoch 44: val_loss did not improve from 4.03318 334/334 [==============================] - 26s 79ms/step - loss: 3.8205 - accuracy: 0.1237 - val_loss: 4.1646 - val_accuracy: 0.1006 Epoch 45/50 334/334 [==============================] - ETA: 0s - loss: 3.8142 - accuracy: 0.1277 Epoch 45: val_loss did not improve from 4.03318 334/334 [==============================] - 26s 78ms/step - loss: 3.8142 - accuracy: 0.1277 - val_loss: 4.0412 - val_accuracy: 0.0731 Epoch 46/50 334/334 [==============================] - ETA: 0s - loss: 3.8040 - accuracy: 0.1305 Epoch 46: val_loss did not improve from 4.03318 334/334 [==============================] - 26s 78ms/step - loss: 3.8040 - accuracy: 0.1305 - val_loss: 4.0406 - val_accuracy: 0.0910 Epoch 47/50 334/334 [==============================] - ETA: 0s - loss: 3.7957 - accuracy: 0.1271 Epoch 47: val_loss improved from 4.03318 to 3.98193, saving model to saved_models\weights.best.from_scratch.hdf5 334/334 [==============================] - 28s 84ms/step - loss: 3.7957 - accuracy: 0.1271 - val_loss: 3.9819 - val_accuracy: 0.1162 Epoch 48/50 334/334 [==============================] - ETA: 0s - loss: 3.7781 - accuracy: 0.1362 Epoch 48: val_loss did not improve from 3.98193 334/334 [==============================] - 39s 115ms/step - loss: 3.7781 - accuracy: 0.1362 - val_loss: 4.0130 - val_accuracy: 0.1030 Epoch 49/50 334/334 [==============================] - ETA: 0s - loss: 3.7635 - accuracy: 0.1356 Epoch 49: val_loss did not improve from 3.98193 334/334 [==============================] - 37s 110ms/step - loss: 3.7635 - accuracy: 0.1356 - val_loss: 4.0001 - val_accuracy: 0.1006 Epoch 50/50 334/334 [==============================] - ETA: 0s - loss: 3.7579 - accuracy: 0.1340 Epoch 50: val_loss did not improve from 3.98193 334/334 [==============================] - 28s 85ms/step - loss: 3.7579 - accuracy: 0.1340 - val_loss: 4.0088 - val_accuracy: 0.1066
def plot_loss_acc(hist, title):
'''plot loss and accuracy of model history object (keras.model.hist)'''
fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(12, 4))
fig.suptitle(title, fontsize=16)
ax1.plot(hist.history["loss"], c="teal", label="loss")
ax1.plot(hist.history["val_loss"], c="teal", label="val_loss", ls="--")
ax1.set_title('Loss')
ax1.legend(loc="upper right")
ax2.plot(hist.history["accuracy"], c="orange", label="accuracy")
ax2.plot(hist.history["val_accuracy"], c="orange", label="val_accuracy", ls="--")
ax2.set_title('Accuracy')
ax2.legend(loc="upper left")
plt.show()
plot_loss_acc(hist, "Dog Breed classifier training (from Scratch)")
model.load_weights('saved_models/weights.best.from_scratch.hdf5')
Try out your model on the test dataset of dog images. Ensure that your test accuracy is greater than 1%.
# get index of predicted dog breed for each image in test set
dog_breed_predictions = [np.argmax(model.predict(np.expand_dims(tensor, axis=0), verbose=0)) for tensor in tqdm(test_tensors)]
# report test accuracy
test_accuracy = 100*np.sum(np.array(dog_breed_predictions)==np.argmax(test_targets, axis=1))/len(dog_breed_predictions)
print('Test accuracy: %.4f%%' % test_accuracy)
100%|████████████████████████████████████████████████████████████████████████████████| 836/836 [00:46<00:00, 17.86it/s]
Test accuracy: 12.2010%
bottleneck_features = np.load('bottleneck_features/DogVGG16Data.npz')
train_VGG16 = bottleneck_features['train']
valid_VGG16 = bottleneck_features['valid']
test_VGG16 = bottleneck_features['test']
The model uses the the pre-trained VGG-16 model as a fixed feature extractor, where the last convolutional output of VGG-16 is fed as input to our model. We only add a global average pooling layer and a fully connected layer, where the latter contains one node for each dog category and is equipped with a softmax.
from keras import layers
VGG16_model = keras.Sequential()
VGG16_model.add(layers.GlobalAveragePooling2D(input_shape=train_VGG16.shape[1:]))
VGG16_model.add(layers.Dense(133, activation='softmax'))
VGG16_model.summary()
Model: "sequential_2"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
global_average_pooling2d_1 (None, 512) 0
(GlobalAveragePooling2D)
dense_1 (Dense) (None, 133) 68229
=================================================================
Total params: 68,229
Trainable params: 68,229
Non-trainable params: 0
_________________________________________________________________
VGG16_model.compile(loss='categorical_crossentropy', optimizer='rmsprop', metrics=['accuracy'])
checkpointer = keras.callbacks.ModelCheckpoint(filepath='saved_models/weights.best.VGG16.hdf5',
verbose=1, save_best_only=True)
VGG16_model.fit(train_VGG16, train_targets,
validation_data=(valid_VGG16, valid_targets),
epochs=20, batch_size=20, callbacks=[checkpointer], verbose=1)
Epoch 1/20 311/334 [==========================>...] - ETA: 0s - loss: 8.1269 - accuracy: 0.2209 Epoch 1: val_loss improved from inf to 3.37334, saving model to saved_models\weights.best.VGG16.hdf5 334/334 [==============================] - 1s 2ms/step - loss: 7.8087 - accuracy: 0.2347 - val_loss: 3.3733 - val_accuracy: 0.4479 Epoch 2/20 314/334 [===========================>..] - ETA: 0s - loss: 2.1499 - accuracy: 0.5930 Epoch 2: val_loss improved from 3.37334 to 2.27155, saving model to saved_models\weights.best.VGG16.hdf5 334/334 [==============================] - 0s 1ms/step - loss: 2.1186 - accuracy: 0.5996 - val_loss: 2.2716 - val_accuracy: 0.6108 Epoch 3/20 322/334 [===========================>..] - ETA: 0s - loss: 1.2170 - accuracy: 0.7410 Epoch 3: val_loss improved from 2.27155 to 2.02929, saving model to saved_models\weights.best.VGG16.hdf5 334/334 [==============================] - 0s 1ms/step - loss: 1.2130 - accuracy: 0.7422 - val_loss: 2.0293 - val_accuracy: 0.6575 Epoch 4/20 316/334 [===========================>..] - ETA: 0s - loss: 0.8122 - accuracy: 0.8136 Epoch 4: val_loss improved from 2.02929 to 1.92801, saving model to saved_models\weights.best.VGG16.hdf5 334/334 [==============================] - 0s 1ms/step - loss: 0.8118 - accuracy: 0.8135 - val_loss: 1.9280 - val_accuracy: 0.6683 Epoch 5/20 321/334 [===========================>..] - ETA: 0s - loss: 0.5381 - accuracy: 0.8612 Epoch 5: val_loss improved from 1.92801 to 1.92219, saving model to saved_models\weights.best.VGG16.hdf5 334/334 [==============================] - 0s 1ms/step - loss: 0.5461 - accuracy: 0.8599 - val_loss: 1.9222 - val_accuracy: 0.6886 Epoch 6/20 319/334 [===========================>..] - ETA: 0s - loss: 0.4038 - accuracy: 0.8937 Epoch 6: val_loss improved from 1.92219 to 1.87971, saving model to saved_models\weights.best.VGG16.hdf5 334/334 [==============================] - 0s 1ms/step - loss: 0.4056 - accuracy: 0.8933 - val_loss: 1.8797 - val_accuracy: 0.6826 Epoch 7/20 321/334 [===========================>..] - ETA: 0s - loss: 0.2939 - accuracy: 0.9179 Epoch 7: val_loss improved from 1.87971 to 1.87124, saving model to saved_models\weights.best.VGG16.hdf5 334/334 [==============================] - 0s 1ms/step - loss: 0.2939 - accuracy: 0.9180 - val_loss: 1.8712 - val_accuracy: 0.7066 Epoch 8/20 316/334 [===========================>..] - ETA: 0s - loss: 0.2227 - accuracy: 0.9366 Epoch 8: val_loss improved from 1.87124 to 1.85954, saving model to saved_models\weights.best.VGG16.hdf5 334/334 [==============================] - 0s 1ms/step - loss: 0.2217 - accuracy: 0.9364 - val_loss: 1.8595 - val_accuracy: 0.7138 Epoch 9/20 300/334 [=========================>....] - ETA: 0s - loss: 0.1709 - accuracy: 0.9482 Epoch 9: val_loss did not improve from 1.85954 334/334 [==============================] - 0s 1ms/step - loss: 0.1706 - accuracy: 0.9484 - val_loss: 1.8749 - val_accuracy: 0.7042 Epoch 10/20 316/334 [===========================>..] - ETA: 0s - loss: 0.1302 - accuracy: 0.9609 Epoch 10: val_loss improved from 1.85954 to 1.80429, saving model to saved_models\weights.best.VGG16.hdf5 334/334 [==============================] - 0s 1ms/step - loss: 0.1330 - accuracy: 0.9600 - val_loss: 1.8043 - val_accuracy: 0.7389 Epoch 11/20 320/334 [===========================>..] - ETA: 0s - loss: 0.0987 - accuracy: 0.9697 Epoch 11: val_loss did not improve from 1.80429 334/334 [==============================] - 0s 1ms/step - loss: 0.1029 - accuracy: 0.9692 - val_loss: 1.8370 - val_accuracy: 0.7305 Epoch 12/20 318/334 [===========================>..] - ETA: 0s - loss: 0.0822 - accuracy: 0.9722 Epoch 12: val_loss did not improve from 1.80429 334/334 [==============================] - 0s 1ms/step - loss: 0.0821 - accuracy: 0.9726 - val_loss: 1.8966 - val_accuracy: 0.7210 Epoch 13/20 321/334 [===========================>..] - ETA: 0s - loss: 0.0618 - accuracy: 0.9829 Epoch 13: val_loss did not improve from 1.80429 334/334 [==============================] - 0s 1ms/step - loss: 0.0613 - accuracy: 0.9828 - val_loss: 1.8671 - val_accuracy: 0.7437 Epoch 14/20 319/334 [===========================>..] - ETA: 0s - loss: 0.0535 - accuracy: 0.9828 Epoch 14: val_loss did not improve from 1.80429 334/334 [==============================] - 0s 1ms/step - loss: 0.0554 - accuracy: 0.9825 - val_loss: 1.8186 - val_accuracy: 0.7413 Epoch 15/20 318/334 [===========================>..] - ETA: 0s - loss: 0.0409 - accuracy: 0.9863 Epoch 15: val_loss did not improve from 1.80429 334/334 [==============================] - 0s 1ms/step - loss: 0.0406 - accuracy: 0.9864 - val_loss: 1.8934 - val_accuracy: 0.7497 Epoch 16/20 320/334 [===========================>..] - ETA: 0s - loss: 0.0270 - accuracy: 0.9917 Epoch 16: val_loss did not improve from 1.80429 334/334 [==============================] - 0s 1ms/step - loss: 0.0276 - accuracy: 0.9916 - val_loss: 1.8691 - val_accuracy: 0.7449 Epoch 17/20 316/334 [===========================>..] - ETA: 0s - loss: 0.0264 - accuracy: 0.9922 Epoch 17: val_loss improved from 1.80429 to 1.77331, saving model to saved_models\weights.best.VGG16.hdf5 334/334 [==============================] - 0s 1ms/step - loss: 0.0255 - accuracy: 0.9927 - val_loss: 1.7733 - val_accuracy: 0.7497 Epoch 18/20 314/334 [===========================>..] - ETA: 0s - loss: 0.0271 - accuracy: 0.9904 Epoch 18: val_loss did not improve from 1.77331 334/334 [==============================] - 0s 1ms/step - loss: 0.0266 - accuracy: 0.9906 - val_loss: 1.9998 - val_accuracy: 0.7449 Epoch 19/20 307/334 [==========================>...] - ETA: 0s - loss: 0.0209 - accuracy: 0.9935 Epoch 19: val_loss did not improve from 1.77331 334/334 [==============================] - 0s 1ms/step - loss: 0.0238 - accuracy: 0.9928 - val_loss: 1.9821 - val_accuracy: 0.7461 Epoch 20/20 318/334 [===========================>..] - ETA: 0s - loss: 0.0172 - accuracy: 0.9945 Epoch 20: val_loss did not improve from 1.77331 334/334 [==============================] - 0s 1ms/step - loss: 0.0183 - accuracy: 0.9946 - val_loss: 2.0145 - val_accuracy: 0.7509
<keras.callbacks.History at 0x1ce70b61c90>
VGG16_model.load_weights('saved_models/weights.best.VGG16.hdf5')
Now, we can use the CNN to test how well it identifies breed within our test dataset of dog images. We print the test accuracy below.
# get index of predicted dog breed for each image in test set
VGG16_predictions = [np.argmax(VGG16_model.predict(np.expand_dims(feature, axis=0), verbose=0)) for feature in tqdm(test_VGG16)]
# report test accuracy
test_accuracy = 100*np.sum(np.array(VGG16_predictions)==np.argmax(test_targets, axis=1))/len(VGG16_predictions)
print('Test accuracy: %.4f%%' % test_accuracy)
100%|████████████████████████████████████████████████████████████████████████████████| 836/836 [00:46<00:00, 18.05it/s]
Test accuracy: 73.5646%
from extract_bottleneck_features import *
def VGG16_predict_breed(img_path):
# extract bottleneck features
bottleneck_feature = extract_VGG16(path_to_tensor(img_path))
# obtain predicted vector
predicted_vector = VGG16_model.predict(bottleneck_feature, verbose=1)
# return dog breed that is predicted by the model
return dog_names[np.argmax(predicted_vector)]
for img in dog_files_short[:1] :
print(img)
display(Image(img))
pred = VGG16_predict_breed(img)
print("Prediction:", pred)
dogImages/train\095.Kuvasz\Kuvasz_06442.jpg
1/1 [==============================] - 0s 192ms/step 1/1 [==============================] - 0s 59ms/step Prediction: Kuvasz
You will now use transfer learning to create a CNN that can identify dog breed from images. Your CNN must attain at least 60% accuracy on the test set.
In Step 4, we used transfer learning to create a CNN using VGG-16 bottleneck features. In this section, you must use the bottleneck features from a different pre-trained model. To make things easier for you, we have pre-computed the features for all of the networks that are currently available in Keras:
The files are encoded as such:
Dog{network}Data.npz
where {network}, in the above filename, can be one of VGG19, Resnet50, InceptionV3, or Xception. Pick one of the above architectures, download the corresponding bottleneck features, and store the downloaded file in the bottleneck_features/ folder in the repository.
In the code block below, extract the bottleneck features corresponding to the train, test, and validation sets by running the following:
bottleneck_features = np.load('bottleneck_features/Dog{network}Data.npz')
train_{network} = bottleneck_features['train']
valid_{network} = bottleneck_features['valid']
test_{network} = bottleneck_features['test']
### TODO: Obtain bottleneck features from another pre-trained CNN.
# These bottleneck files where downloaded separately and stored in bottleneck_features/
# https://s3-us-west-1.amazonaws.com/udacity-aind/dog-project/DogVGG19Data.npz
# https://s3-us-west-1.amazonaws.com/udacity-aind/dog-project/DogResnet50Data.npz
# https://s3-us-west-1.amazonaws.com/udacity-aind/dog-project/DogInceptionV3Data.npz
# https://s3-us-west-1.amazonaws.com/udacity-aind/dog-project/DogXceptionData.npz
# load bottleneck features for all four pretrained networks
networks_pretrained= []
for network in ["VGG19", "ResNet50", "Xception", "InceptionV3"]:
bnf_path = 'bottleneck_features/Dog{}Data.npz'.format(network)
print("{:15} loading {} ...".format(network, bnf_path))
bottleneck_features = np.load(bnf_path)
d = {
"network": network,
"train": bottleneck_features['train'],
"valid": bottleneck_features['valid'],
"test": bottleneck_features['test']
}
networks_pretrained.append(d)
VGG19 loading bottleneck_features/DogVGG19Data.npz ... ResNet50 loading bottleneck_features/DogResNet50Data.npz ... Xception loading bottleneck_features/DogXceptionData.npz ... InceptionV3 loading bottleneck_features/DogInceptionV3Data.npz ...
# Loop : for each network, add the same basic layers, compile and fit
for n in networks_pretrained :
name = n["network"]
model = keras.Sequential(name = name)
input_shape = n["train"].shape[1:]
model.add(layers.GlobalAveragePooling2D(input_shape=input_shape))
model.add(layers.Dense(133, activation='softmax'))
model.summary()
print("{:15} compiling ...".format(name))
model.compile(loss='categorical_crossentropy', optimizer='rmsprop', metrics=['accuracy'])
print("{:15} training model ...".format(name))
checkpointer = keras.callbacks.ModelCheckpoint(
filepath='saved_models/weights.best.{}.hdf5'.format(name),
verbose=0,
save_best_only=True
)
n["model"] = model
# fit for 30 epochs
n["hist"] = model.fit(
n["train"],
train_targets,
validation_data=(n["valid"], valid_targets),
epochs=20,
batch_size=20,
callbacks=[checkpointer],
verbose=0)
# print for insight
plot_loss_acc(n["hist"], n["network"])
Model: "VGG19"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
global_average_pooling2d_3 (None, 512) 0
(GlobalAveragePooling2D)
dense_3 (Dense) (None, 133) 68229
=================================================================
Total params: 68,229
Trainable params: 68,229
Non-trainable params: 0
_________________________________________________________________
VGG19 compiling ...
VGG19 training model ...
Model: "ResNet50"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
global_average_pooling2d_4 (None, 2048) 0
(GlobalAveragePooling2D)
dense_4 (Dense) (None, 133) 272517
=================================================================
Total params: 272,517
Trainable params: 272,517
Non-trainable params: 0
_________________________________________________________________
ResNet50 compiling ...
ResNet50 training model ...
Model: "Xception"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
global_average_pooling2d_5 (None, 2048) 0
(GlobalAveragePooling2D)
dense_5 (Dense) (None, 133) 272517
=================================================================
Total params: 272,517
Trainable params: 272,517
Non-trainable params: 0
_________________________________________________________________
Xception compiling ...
Xception training model ...
Model: "InceptionV3"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
global_average_pooling2d_6 (None, 2048) 0
(GlobalAveragePooling2D)
dense_6 (Dense) (None, 133) 272517
=================================================================
Total params: 272,517
Trainable params: 272,517
Non-trainable params: 0
_________________________________________________________________
InceptionV3 compiling ...
InceptionV3 training model ...
# Loop : test all predictions and compare
for n in networks_pretrained :
# load best weights from file
n["model"].load_weights('saved_models/weights.best.{}.hdf5'.format(n["network"]))
print("{:15} Testing ...".format(n["network"]))
# get index of predicted dog breed for each image in test set
predictions = [np.argmax(n["model"].predict(np.expand_dims(feature, axis=0), verbose=0)) for feature in tqdm(n["test"])]
# report test accuracy
test_accuracy = 100*np.sum(np.array(predictions)==np.argmax(test_targets, axis=1))/len(predictions)
print(n["network"], 'Test accuracy: %.4f%%' % test_accuracy)
VGG19 Testing ...
100%|████████████████████████████████████████████████████████████████████████████████| 836/836 [00:46<00:00, 18.02it/s]
VGG19 Test accuracy: 72.2488% ResNet50 Testing ...
100%|████████████████████████████████████████████████████████████████████████████████| 836/836 [00:46<00:00, 17.94it/s]
ResNet50 Test accuracy: 81.1005% Xception Testing ...
100%|████████████████████████████████████████████████████████████████████████████████| 836/836 [00:46<00:00, 18.10it/s]
Xception Test accuracy: 84.2105% InceptionV3 Testing ...
100%|████████████████████████████████████████████████████████████████████████████████| 836/836 [00:46<00:00, 18.07it/s]
InceptionV3 Test accuracy: 79.1866%
Create a CNN to classify dog breed. At the end of your code cell block, summarize the layers of your model by executing the line:
<your model's name>.summary()
Question 5: Outline the steps you took to get to your final CNN architecture and your reasoning at each step. Describe why you think the architecture is suitable for the current problem.
Answer:
### TODO: Define your architecture.
# get bottleneck features from Xception from file
Xception_bottleneck_features = np.load('bottleneck_features/DogXceptionData.npz')
Xception_train = Xception_bottleneck_features['train']
Xception_valid = Xception_bottleneck_features['valid']
Xception_test = Xception_bottleneck_features['test']
# # Configure input and add layers :
# # Note this one gives acc 84% after 3-4 epochs
model = keras.Sequential(name = "Xception-final")
# add a global spatial average pooling layer.
# add a fully-connected layer designed to distinguish between the 133 breeds.
model.add(layers.GlobalAveragePooling2D(input_shape=Xception_train.shape[1:]))
model.add(layers.Dense(133, activation='softmax'))
# Print summary
model.summary()
Model: "Xception-final"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
global_average_pooling2d_7 (None, 2048) 0
(GlobalAveragePooling2D)
dense_7 (Dense) (None, 133) 272517
=================================================================
Total params: 272,517
Trainable params: 272,517
Non-trainable params: 0
_________________________________________________________________
### TODO: Compile the model.
model.compile(loss='categorical_crossentropy', optimizer='rmsprop', metrics=['accuracy'])
# model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
Train your model in the code cell below. Use model checkpointing to save the model that attains the best validation loss.
You are welcome to augment the training data, but this is not a requirement.
# ### TODO: Train the model.
print(model.name)
checkpointer = keras.callbacks.ModelCheckpoint(
filepath='saved_models/weights.best.{}.hdf5'.format(model.name),
verbose=1,
save_best_only=True
)
hist = model.fit(
Xception_train, train_targets,
validation_data=(
Xception_valid,
valid_targets),
epochs=10,
batch_size=20,
callbacks=[checkpointer],
verbose=1
)
# plot loss and accuracy
plot_loss_acc(hist, model.name)
Xception-final Epoch 1/10 324/334 [============================>.] - ETA: 0s - loss: 1.0726 - accuracy: 0.7341 Epoch 1: val_loss improved from inf to 0.51898, saving model to saved_models\weights.best.Xception-final.hdf5 334/334 [==============================] - 2s 4ms/step - loss: 1.0573 - accuracy: 0.7362 - val_loss: 0.5190 - val_accuracy: 0.8311 Epoch 2/10 319/334 [===========================>..] - ETA: 0s - loss: 0.3997 - accuracy: 0.8730 Epoch 2: val_loss improved from 0.51898 to 0.49622, saving model to saved_models\weights.best.Xception-final.hdf5 334/334 [==============================] - 2s 5ms/step - loss: 0.3974 - accuracy: 0.8738 - val_loss: 0.4962 - val_accuracy: 0.8395 Epoch 3/10 328/334 [============================>.] - ETA: 0s - loss: 0.3205 - accuracy: 0.9026 Epoch 3: val_loss did not improve from 0.49622 334/334 [==============================] - 1s 4ms/step - loss: 0.3200 - accuracy: 0.9028 - val_loss: 0.5127 - val_accuracy: 0.8515 Epoch 4/10 332/334 [============================>.] - ETA: 0s - loss: 0.2737 - accuracy: 0.9102 Epoch 4: val_loss did not improve from 0.49622 334/334 [==============================] - 1s 3ms/step - loss: 0.2754 - accuracy: 0.9103 - val_loss: 0.4980 - val_accuracy: 0.8515 Epoch 5/10 325/334 [============================>.] - ETA: 0s - loss: 0.2482 - accuracy: 0.9248 Epoch 5: val_loss did not improve from 0.49622 334/334 [==============================] - 1s 3ms/step - loss: 0.2447 - accuracy: 0.9256 - val_loss: 0.5360 - val_accuracy: 0.8599 Epoch 6/10 331/334 [============================>.] - ETA: 0s - loss: 0.2228 - accuracy: 0.9310 Epoch 6: val_loss did not improve from 0.49622 334/334 [==============================] - 1s 4ms/step - loss: 0.2218 - accuracy: 0.9311 - val_loss: 0.5474 - val_accuracy: 0.8527 Epoch 7/10 332/334 [============================>.] - ETA: 0s - loss: 0.1918 - accuracy: 0.9410 Epoch 7: val_loss did not improve from 0.49622 334/334 [==============================] - 1s 3ms/step - loss: 0.1926 - accuracy: 0.9407 - val_loss: 0.5479 - val_accuracy: 0.8491 Epoch 8/10 327/334 [============================>.] - ETA: 0s - loss: 0.1775 - accuracy: 0.9465 Epoch 8: val_loss did not improve from 0.49622 334/334 [==============================] - 1s 3ms/step - loss: 0.1764 - accuracy: 0.9469 - val_loss: 0.5572 - val_accuracy: 0.8503 Epoch 9/10 325/334 [============================>.] - ETA: 0s - loss: 0.1563 - accuracy: 0.9517 Epoch 9: val_loss did not improve from 0.49622 334/334 [==============================] - 1s 3ms/step - loss: 0.1564 - accuracy: 0.9513 - val_loss: 0.5495 - val_accuracy: 0.8539 Epoch 10/10 329/334 [============================>.] - ETA: 0s - loss: 0.1436 - accuracy: 0.9555 Epoch 10: val_loss did not improve from 0.49622 334/334 [==============================] - 2s 5ms/step - loss: 0.1432 - accuracy: 0.9555 - val_loss: 0.5717 - val_accuracy: 0.8527
### TODO: Load the model weights with the best validation loss.
model.load_weights('saved_models/weights.best.Xception-final.hdf5')
Try out your model on the test dataset of dog images. Ensure that your test accuracy is greater than 60%.
### TODO: Calculate classification accuracy on the test dataset.
# get index of predicted dog breed for each image in test set
predictions = [np.argmax(model.predict(np.expand_dims(feature, axis=0), verbose=0)) for feature in tqdm(Xception_test)]
# report test accuracy
test_accuracy = 100*np.sum(np.array(predictions)==np.argmax(test_targets, axis=1))/len(predictions)
print(model.name, ' test accuracy: %.4f%%' % test_accuracy)
100%|████████████████████████████████████████████████████████████████████████████████| 836/836 [00:33<00:00, 25.12it/s]
Xception-final test accuracy: 0.0000%
Write a function that takes an image path as input and returns the dog breed (Affenpinscher, Afghan_hound, etc) that is predicted by your model.
Similar to the analogous function in Step 5, your function should have three steps:
dog_names array defined in Step 0 of this notebook to return the corresponding breed.The functions to extract the bottleneck features can be found in extract_bottleneck_features.py, and they have been imported in an earlier code cell. To obtain the bottleneck features corresponding to your chosen CNN architecture, you need to use the function
extract_{network}
where {network}, in the above filename, should be one of VGG19, Resnet50, InceptionV3, or Xception.
### TODO: Write a function that takes a path to an image as input
### and returns the dog breed that is predicted by the model.
print(model.name, " predictions...")
def final_model_predict_breed(img_path):
'''Extract features using built-in keras Xception module
Then do the prediction with our model and return dog breed using
dog_names labels.'''
# load labels
dog_names = [item[20:-1] for item in sorted(glob("dogImages/train/*/"))]
# load feature extraction model Xception
bottleneck_feature = extract_Xception(path_to_tensor(img_path))
# load weights and make predictions
model.load_weights('saved_models/weights.best.Xception-final.hdf5')
predicted_vector = model.predict(bottleneck_feature, verbose=0)
return dog_names[np.argmax(predicted_vector)]
# run and show some predictions :)
dog_files_mini = dog_files_short[:10]
preds = [final_model_predict_breed(img) for img in dog_files_mini]
fig_labels = ['file: {}<br>prediction: {}'.format(f.split('\\')[-1],p) for f,p in zip(dog_files_mini, preds)]
gallery(dog_files_mini, fig_labels)
Xception-final predictions... 1/1 [==============================] - 1s 625ms/step 1/1 [==============================] - 1s 901ms/step 1/1 [==============================] - 1s 636ms/step 1/1 [==============================] - 1s 624ms/step 1/1 [==============================] - 1s 627ms/step 1/1 [==============================] - 1s 627ms/step 1/1 [==============================] - 1s 630ms/step 1/1 [==============================] - 1s 627ms/step 1/1 [==============================] - 1s 638ms/step 1/1 [==============================] - 1s 629ms/step
Write an algorithm that accepts a file path to an image and first determines whether the image contains a human, dog, or neither. Then,
You are welcome to write your own functions for detecting humans and dogs in images, but feel free to use the face_detector and dog_detector functions developed above. You are required to use your CNN from Step 5 to predict dog breed.
Some sample output for our algorithm is provided below, but feel free to design your own user experience!

### TODO: Write your algorithm.
### Feel free to use as many code cells as needed.
def find_human(img_arr, algorithm="dlib-cnn"):
'''Uses CNN-MMOD algorithm to detect human faces in img_arr.
Returns face position if a human face is found, otherwise returns False'''
import cv2
import imutils
from dlib import cnn_face_detection_model_v1
if algorithm == "dlib-cnn":
# more accurate, but slow, this detector is used by default
# print('using dlib cnn model for human detection...')
face_cnn = cnn_face_detection_model_v1("mmod/mmod_human_face_detector.dat")
rects = face_cnn(img_arr, 1)
if len(rects) > 0:
# we are only getting one human face (rects[0])
face = imutils.face_utils.rect_to_bb(rects[0].rect)
# face = tuple([int(resize_ratio*x) for x in face])
confidence = rects[0].confidence
return face, confidence
else:
return None
if algorithm == "hog" :
# algorithm HoG, this one is faster but has higher error.
# print('using hog-svm for human detection...')
gray = gray = cv2.cvtColor(img_arr, cv2.COLOR_BGR2GRAY)
face_hog = dlib.get_frontal_face_detector()
rects = face_hog(gray, 1)
if rects :
face = imutils.face_utils.rect_to_bb(rects[0])
confidence = 1.0
return face, confidence
else:
return None
def ResNet50_predict_labels(img_path):
'''Returns prediction vector for image located at img_path.'''
from keras.applications.resnet import ResNet50
from keras.applications.resnet import preprocess_input
from keras.applications.resnet import decode_predictions
img = preprocess_input(path_to_tensor(img_path))
return np.argmax(ResNet50_model.predict(img, verbose=0))
def find_dog(img_path):
'''Returns True if a dog is detected in the image stored at img_path.'''
prediction = ResNet50_predict_labels(img_path)
return ((prediction <= 268) & (prediction >= 151))
def find_dogbreed(img_path):
'''Extract features using built-in keras Xception module
Then do the prediction with our model and return dog breed using
dog_names labels.'''
# load labels
dog_names = [item[20:-1] for item in sorted(glob("dogImages/train/*/"))]
# load feature extraction model Xception
bottleneck_feature = extract_Xception(path_to_tensor(img_path))
# load weights and make predictions
model.load_weights('saved_models/weights.best.Xception-final.hdf5')
predicted_vector = model.predict(bottleneck_feature, verbose=0)
return dog_names[np.argmax(predicted_vector)]
def draw_bow(face, img_arr, text, color):
'''draw boxes using faces coordinates and annotates text in image'''
x,y,w,h = face
cv2.rectangle(
img_arr,
(x,y), (x+w,y+h),
color=color,
thickness=2,
lineType=cv2.LINE_AA
)
cv2.putText(
img_arr,
text,
(x,y-3),
fontFace=cv2.FONT_HERSHEY_DUPLEX,
fontScale=0.7,
color=color,
lineType=cv2.LINE_AA
)
return img_arr
def predict_final(img_path):
'''Accepts a file path to an image and first determines whether
the image contains a human, dog, or neither.
- if a dog is detected, return the predicted breed.
- if a human is detected, return the resembling dog breed.
- if neither is detected, provide output that indicates an error.
'''
import time
print("Analyzing image: {}".format(img_path))
start = time.time()
IMG_MAX_WIDTH = 600
HUMAN_COLOR = (0,0,255)
is_dog = False
breed = None #used for both dogs and humans lol
is_human = False
human_pos = None
# import to np array and resize if size is exceeded.
img_arr = cv2.imread(img_path)
if img_arr.shape[0] > IMG_MAX_WIDTH:
img_arr = imutils.resize(img_arr, width=IMG_MAX_WIDTH)
# look for dog and humans
is_dog = find_dog(img_path)
human_pos = find_human(img_arr, algorithm="hog")
is_human = human_pos is not None
# predict breed
if is_dog:
print("Doggo 🐕 found !")
breed = find_dogbreed(img_path)
print("Looks like a {}.".format(breed))
elif is_human and not is_dog:
print("Human 👩👨detected !")
visual = draw_bow(
human_pos[0], img_arr,
text='human !?', color=HUMAN_COLOR
)
breed = find_dogbreed(img_path)
print("Anyway, he/she looks like a {}.".format(breed))
elif is_dog and is_human:
print("❌ Not sure if I see a a dog or a human!, predicting breed anyway !")
else :
# not human not dog
print("❌ I wasn't able to find any faces in the picture :'(")
# convert BGR image to RGB
img_arr = cv2.cvtColor(img_arr, cv2.COLOR_BGR2RGB)
# print("elapsed time : {}".format(time.time() - start))
results = {
"is_dog" : is_dog,
"breed" : breed,
"is_human" : is_human,
"human_pos" : human_pos,
}
return (results, img_arr)
In this section, you will take your new algorithm for a spin! What kind of dog does the algorithm think that you look like? If you have a dog, does it predict your dog's breed accurately? If you have a cat, does it mistakenly think that your cat is a dog?
Test your algorithm at least six images on your computer. Feel free to use any images you like. Use at least two human and two dog images.
Question 6: Is the output better than you expected :) ? Or worse :( ? Provide at least three possible points of improvement for your algorithm.
Answer:
## TODO: Execute your algorithm from Step 6 on
## at least 6 images on your computer.
## Feel free to use as many code cells as needed.
from glob import glob
from pprint import pprint
import matplotlib.image as mpimg
test_images = glob("random_test_images/*/*")
for i in test_images:
res = predict_final(i)
pprint(res[0])
# display the image
plt.imshow(res[1])
plt.axis("off")
plt.show()
# print some images from the training set
if res[0]["breed"] != None :
qty = 7
example_dir = [dir for dir in glob("dogImages/train/*/") if res[0]["breed"] == dir[20:-1]][0]
fig, ax = plt.subplots(1, qty, figsize=(12, 2))
fig.suptitle("Other {} examples".format(res[0]['breed']))
for i, ex in enumerate(glob(example_dir+"/*")[:qty]):
ax[i].imshow(mpimg.imread(ex))
ax[i].axis("off")
plt.show()
print("End\n\n")
Analyzing image: random_test_images\dogs\0002-inmortalidadcangrejo (Small).png
Doggo 🐕 found !
1/1 [==============================] - 1s 860ms/step
Looks like a Icelandic_sheepdog.
{'breed': 'Icelandic_sheepdog',
'human_pos': None,
'is_dog': True,
'is_human': False}
End
Analyzing image: random_test_images\dogs\001.jpeg
Doggo 🐕 found !
1/1 [==============================] - 1s 629ms/step
Looks like a Collie.
{'breed': 'Collie', 'human_pos': None, 'is_dog': True, 'is_human': False}
End
Analyzing image: random_test_images\dogs\1355996702_A-brown-and-white-Boxane-dog.jpg
Doggo 🐕 found !
1/1 [==============================] - 1s 630ms/step
Looks like a Boxer.
{'breed': 'Boxer', 'human_pos': None, 'is_dog': True, 'is_human': False}
End
Analyzing image: random_test_images\dogs\234023992e6c4813861628774eabafd0.jpg
Doggo 🐕 found !
1/1 [==============================] - 1s 629ms/step
Looks like a Neapolitan_mastiff.
{'breed': 'Neapolitan_mastiff',
'human_pos': None,
'is_dog': True,
'is_human': False}
End
Analyzing image: random_test_images\dogs\246145-425x340-schnoodle-dog-cute.jpg
Doggo 🐕 found !
1/1 [==============================] - 1s 635ms/step
Looks like a Miniature_schnauzer.
{'breed': 'Miniature_schnauzer',
'human_pos': None,
'is_dog': True,
'is_human': False}
End
Analyzing image: random_test_images\dogs\5b9eb81a-6e08-4c11-9ded-885e073a975d.jpg
Doggo 🐕 found !
1/1 [==============================] - 1s 626ms/step
Looks like a Norwegian_buhund.
{'breed': 'Norwegian_buhund',
'human_pos': None,
'is_dog': True,
'is_human': False}
End
Analyzing image: random_test_images\dogs\5cd4358d7fc0a.webp
Doggo 🐕 found !
1/1 [==============================] - 1s 629ms/step
Looks like a Chihuahua.
{'breed': 'Chihuahua', 'human_pos': None, 'is_dog': True, 'is_human': False}
End
Analyzing image: random_test_images\dogs\9927eb4c-c9ab-4735-9308-95a108a21d44.jpg
Doggo 🐕 found !
1/1 [==============================] - 1s 908ms/step
Looks like a Canaan_dog.
{'breed': 'Canaan_dog', 'human_pos': None, 'is_dog': True, 'is_human': False}
End
Analyzing image: random_test_images\dogs\american_staff.webp
Doggo 🐕 found !
1/1 [==============================] - 1s 631ms/step
Looks like a American_staffordshire_terrier.
{'breed': 'American_staffordshire_terrier',
'human_pos': None,
'is_dog': True,
'is_human': False}
End
Analyzing image: random_test_images\dogs\black-labrodor-retriever.jpg
Doggo 🐕 found !
1/1 [==============================] - 1s 627ms/step
Looks like a Labrador_retriever.
{'breed': 'Labrador_retriever',
'human_pos': None,
'is_dog': True,
'is_human': False}
End
Analyzing image: random_test_images\dogs\Canaan-Dog-294x300.jpg
Doggo 🐕 found !
1/1 [==============================] - 1s 626ms/step
Looks like a Basenji.
{'breed': 'Basenji', 'human_pos': None, 'is_dog': True, 'is_human': False}
End
Analyzing image: random_test_images\dogs\DachshundPurebredDogWillowStandard3YearsOld4.jpg
Doggo 🐕 found !
1/1 [==============================] - 1s 625ms/step
Looks like a Dachshund.
{'breed': 'Dachshund', 'human_pos': None, 'is_dog': True, 'is_human': False}
End
Analyzing image: random_test_images\dogs\Dallas_edit2-1024x745.jpg
Doggo 🐕 found !
1/1 [==============================] - 1s 898ms/step
Looks like a Dachshund.
{'breed': 'Dachshund', 'human_pos': None, 'is_dog': True, 'is_human': False}
End
Analyzing image: random_test_images\dogs\Dog-Sitting.jpg
Doggo 🐕 found !
1/1 [==============================] - 1s 633ms/step
Looks like a Great_pyrenees.
{'breed': 'Great_pyrenees',
'human_pos': None,
'is_dog': True,
'is_human': False}
End
Analyzing image: random_test_images\dogs\fdaeea93-a12c-446f-95d7-d7251c491fc9.jpg
Doggo 🐕 found !
1/1 [==============================] - 1s 610ms/step
Looks like a Bouvier_des_flandres.
{'breed': 'Bouvier_des_flandres',
'human_pos': None,
'is_dog': True,
'is_human': False}
End
Analyzing image: random_test_images\dogs\Feature-Curly-Haired-Dog-Breeds-scaled.jpg
Doggo 🐕 found !
1/1 [==============================] - 1s 635ms/step
Looks like a Lakeland_terrier.
{'breed': 'Lakeland_terrier',
'human_pos': None,
'is_dog': True,
'is_human': False}
End
Analyzing image: random_test_images\dogs\LaizhouHongRedDogBambooTailKennelDog2.jpg
Doggo 🐕 found !
1/1 [==============================] - 1s 631ms/step
Looks like a Doberman_pinscher.
{'breed': 'Doberman_pinscher',
'human_pos': None,
'is_dog': True,
'is_human': False}
End
Analyzing image: random_test_images\dogs\Norwegian-Puffin-Dog.jpg
Doggo 🐕 found !
1/1 [==============================] - 1s 632ms/step
Looks like a Norwegian_lundehund.
{'breed': 'Norwegian_lundehund',
'human_pos': None,
'is_dog': True,
'is_human': False}
End
Analyzing image: random_test_images\dogs\pug-fattest-dogs-1600394338.jpg
Doggo 🐕 found !
1/1 [==============================] - 1s 631ms/step
Looks like a Bulldog.
{'breed': 'Bulldog', 'human_pos': None, 'is_dog': True, 'is_human': False}
End
Analyzing image: random_test_images\dogs\ReagleRottweilerBeagleMixedBreedDogAtariPuppy4Months1.jpg
Doggo 🐕 found !
1/1 [==============================] - 1s 902ms/step
Looks like a Entlebucher_mountain_dog.
{'breed': 'Entlebucher_mountain_dog',
'human_pos': None,
'is_dog': True,
'is_human': False}
End
Analyzing image: random_test_images\dogs\trpd_k9_boris-731x1024.jpg
Doggo 🐕 found !
1/1 [==============================] - 1s 630ms/step
Looks like a Norwegian_elkhound.
{'breed': 'Norwegian_elkhound',
'human_pos': None,
'is_dog': True,
'is_human': False}
End
Analyzing image: random_test_images\dogs\whippet-dog_fvy2bw.webp
Doggo 🐕 found !
1/1 [==============================] - 1s 615ms/step
Looks like a Italian_greyhound.
{'breed': 'Italian_greyhound',
'human_pos': None,
'is_dog': True,
'is_human': False}
End
Analyzing image: random_test_images\dogs\yorkshire-terrier-running-on-lawn-520254142-5ab57cf1c06471003620f804.jpg
Doggo 🐕 found !
1/1 [==============================] - 1s 633ms/step
Looks like a Norwich_terrier.
{'breed': 'Norwich_terrier',
'human_pos': None,
'is_dog': True,
'is_human': False}
End
Analyzing image: random_test_images\humans\02-snoop-dogg-2017-cant-stop-wont-stop-premier-red-carpet-billboard-1548-compressed.jpg
Human 👩👨detected !
1/1 [==============================] - 1s 632ms/step
Anyway, he/she looks like a Smooth_fox_terrier.
{'breed': 'Smooth_fox_terrier',
'human_pos': ((180, 57, 186, 185), 1.0),
'is_dog': False,
'is_human': True}
End
Analyzing image: random_test_images\humans\human-face-4.webp
Human 👩👨detected !
1/1 [==============================] - 1s 909ms/step
Anyway, he/she looks like a Bichon_frise.
{'breed': 'Bichon_frise',
'human_pos': ((236, 133, 154, 155), 1.0),
'is_dog': False,
'is_human': True}
End
Analyzing image: random_test_images\humans\th-2513149514.jpg
Human 👩👨detected !
1/1 [==============================] - 1s 625ms/step
Anyway, he/she looks like a Anatolian_shepherd_dog.
{'breed': 'Anatolian_shepherd_dog',
'human_pos': ((103, 68, 108, 107), 1.0),
'is_dog': False,
'is_human': True}
End
Analyzing image: random_test_images\humans\th.webp
Human 👩👨detected !
1/1 [==============================] - 1s 629ms/step
Anyway, he/she looks like a Dachshund.
{'breed': 'Dachshund',
'human_pos': ((44, 80, 107, 107), 1.0),
'is_dog': False,
'is_human': True}
End